Message-ID: <TYWP286MB226710E14843B461637C4DC6B9719@TYWP286MB2267.JPNP286.PROD.OUTLOOK.COM>
Date:   Fri, 10 Dec 2021 01:07:26 +0000
From:   小川　修一 
        <ogawa.syuuichi@...d-mse.com>
To:     "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>
CC:     山本　達史 
        <yamamoto.tatsushi@...d-mse.com>,
        "Natsume, Wataru (ADITJ/SWG)" <wnatsume@...adit-jv.com>,
        小川　修一 
        <ogawa.syuuichi@...d-mse.com>
Subject: RE: invalid counter value in request_queue

I forgot to take care gdb-script about ATOMIC mode in percpu_count_ptr.   I fixed the script then I got:

super_block 0xffff9bb3f8bbcc30 "vdb"
 q=(struct request_queue *) 0xffff9bb3fb53d1c0,q->q_usage_counter.percpu_count_ptr=(unsigned long *) 0x3a80c000b2b9
 0:0xffffd634bfc0b2b8,0x0,0
 1:0xffffd634bfc8b2b8,0x0,0
 2:0xffffd634bfd0b2b8,0x0,0

Result is as same as previous my e-mail.

-----Original Message-----
From: 小川　修一 
Sent: Friday, December 10, 2021 9:54 AM
To: linux-kernel@...r.kernel.org
Cc: 山本　達史 <yamamoto.tatsushi@...d-mse.com>; Natsume, Wataru (ADITJ/SWG) <wnatsume@...adit-jv.com>
Subject: RE: invalid counter value in request_queue

Hi all

I understood the reason why per-cpu couner became negative.
As I mention previous mail, q_usage_counter should be atomic counter.
percpu_refcount has not only percpu counter but also ATOMIC counter mode.

As fact, blk_alloc_queue_node() initialize request_queue->q_usage_counter as ATOMIC_MODE like this:

        /*
         * Init percpu_ref in atomic mode so that it's faster to shutdown.
         * See blk_register_queue() for details.
         */
        if (percpu_ref_init(&q->q_usage_counter,
                                blk_queue_usage_counter_release,
                                PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))

However, q_usage_counter becomes percpu mode at blk_register_queue() .

blk_register_queue()

        /*
         * SCSI probing may synchronously create and destroy a lot of
         * request_queues for non-existent devices.  Shutting down a fully
         * functional queue takes measureable wallclock time as RCU grace
         * periods are involved.  To avoid excessive latency in these
         * cases, a request_queue starts out in a degraded mode which is
         * faster to shut down and is made fully functional here as
         * request_queues for non-existent devices never get registered.
         */
        if (!blk_queue_init_done(q)) {
                queue_flag_set_unlocked(QUEUE_FLAG_INIT_DONE, q);
                percpu_ref_switch_to_percpu(&q->q_usage_counter);
                blk_queue_bypass_end(q);
        }


When I remove percpu_ref_switch_to_percpu(&q->q_usage_counter) in blk_register_queue() ;

- At least percpu counter is no longer used.

super_block 0xffff9bb3f8bbcc30 "vdb"
 q=(struct request_queue *) 0xffff9bb3fb53d1c0,q->q_usage_counter.percpu_count_ptr=(unsigned long *) 0x3a80c000b2b9
 0:0xffffd634bfc0b2b9,0x0,0
 1:0xffffd634bfc8b2b9,0x0,0
 2:0xffffd634bfd0b2b9,0x0,0

You can find percpu_count_ptr has a bit __PERCPU_REF_ATOMIC=0x1 per cpu counter looks not used

- I check again q_usage_counter.

crash> p $q->q_usage_counter
$1 = {
  count = {
    counter = 5
  },
  percpu_count_ptr = 64324651496121,
  release = 0xffffffff9d44445d,
  confirm_switch = 0x0,
  force_atomic = true,
  rcu = {
    next = 0x0,
    func = 0x0
  }
}

atomic_t count is 5.
5 is OK or NG ? I will check it.
Other block device which ext4 mount on has counter=1

Now question is 
   Why blk_register_queue() turn q_usage_counter to percpu mode ?

I think this code has some reason to turn to percpu mode, though blk-core.c treat that counter as atomic counter.
Does it have some problem ?

Best Regards, shuichi ogawa 

-----Original Message-----
From: 小川　修一 <ogawa.syuuichi@...d-mse.com>
Sent: Thursday, December 9, 2021 1:50 PM
To: linux-kernel@...r.kernel.org
Cc: 小川　修一 <ogawa.syuuichi@...d-mse.com>; 山本　達史 <yamamoto.tatsushi@...d-mse.com>; Natsume, Wataru (ADITJ/SWG) <wnatsume@...adit-jv.com>
Subject: invalid counter value in request_queue

Hi, all
I have first time to post mail, so if you have some matter, please let me know.
I'm studying Linux kernel, referencing kdump, to clarify system performance problem.
Now I found strange value in request_queue->q_usage_counter.percpu_count_ptr

Kernel version: 4.9.52, I checked 5.10.80 briefly, and looked similar.

super_block 0xffff9a563820e430 "vdb"
 q=(struct request_queue *) 0xffff9a563b948920,q->q_usage_counter.percpu_count_ptr=(unsigned long *) 0x39dbc000b2b8
 0:0xffffd431ffc0b2b8,0xffffffffffffdaf1,-9487
 1:0xffffd431ffc8b2b8,0x0,0
 2:0xffffd431ffd0b2b8,0x2510,9488

This is output of gdb script in crash commadn. Format is <cpu>:<address>,<value in HEX>, <value in signed long DEC>

Values of percpu_counter_ptr, big value or negative value on cpu0, and positive value on cpu2.
If sum up all cpus, total=1, it means 1 request remain in /dev/vdb at that kdump.

Easy to estimate, count up cpu and count down cpu are different.
I think the q_usage_counter doesn't work as reference counter to guard invalid disposing request queue, however I don't found to use this counter.

System looks no problem. However I wonder that causes any troubles like invalid disposing resource.
I ask you that this is really no problem at all.

---

As we know, this counter is reference counter of request queue access, For example generic_make_request
   blk_queue_enter(q, false) -> percpu_ref_tryget_live(&q->q_usage_counter) : count up
   blk_queue_exit(q) -> percpu_ref_put(&q->q_usage_counter) : count down

If count up on cpu2, and count down on cpu0, this counter becomes invalid.
I found 2 cases:

case-1: normal case of counting actual requested I/O

blk_mq_map_request() request bio to block device, then count up in blk_queue_enter_live(q)
blk_mq_end_request() called at terminating I/O at IRQ context, then count down in
  blk_mq_free_request() -> blk_queue_exit(q)

IRQ context is normally run on cpu0 in my system. so If AP requests FILE-I/O on cpu2, this problem is reproduced.

case-2: preemption

generic_make_request is not preempt disabled, then cpu may changes between blk_queue_enter and blk_queue_exit.

Now I think q_usage_counter should consist of simple atomic_t or kref_t instead of percpu_ref.
System looks no problem as of now, I've not yet make any patches to correct it.
If I have a chance to make the patch, I will post again.

Best Regards, shuichi ogawa, NTT-Data MSE corporation

Powered by blists - more mailing lists